MACF-Net: Multi-Level Aware Cross-Modal Fusion Network for 3D Object Detection

Detection of small or distant objects in large-scale scenes remains one of the main challenges for high-precision 3D object detection in autonomous driving. Although multiple sensor fusion has become increasingly common for this task, existing fusion methods still struggle with issues like occlusion and weak feature representation of small or distant objects. To this end, we propose MACF-Net, a cross-modal fusion 3D object detection network suitable for small or distant object. Specifically, we propose an image-guided dynamic sampling strategy to enhance point density in distant regions. We further integrate feature information from images, point clouds coupled with voxels by achieving cross-modal alignment through geometric projection. By employing voxel-based and point-based fusion modules, we achieve cross-modal feature fusion at both voxel and point levels, effectively leveraging the texture details from images and the geometric depth from point clouds. Finally, a multi-feature fused module aggregates the multi-level fused features to produce refined bounding box predictions and confidence scores. The experimental results on the KITTI dataset demonstrate that MACF-Net outperforms existing state-of-the-art methods, highlighting its effectiveness in the detection of small or distant objects.